Silver paste conductivity data modeling and formula optimizing system based on machine learning

Abstract

The invention relates to the technical field of silver paste preparation, in particular to a silver paste conductivity data modeling and formula optimization system based on machine learning, which comprises a data acquisition and storage module, a preprocessing module, a characteristic influence analysis module, a formula optimization module, a simulation verification module and the like. The method comprises the following steps: acquiring original data of a silver paste formula, preprocessing, calculating influence coefficients of all components on target performance by utilizing a machine learning model, and identifying high and low influence components; the formula optimization module is combined with component content constraints and adopts a multi-objective optimization algorithm to generate candidate formulas; and the simulation verification module verifies the performance of the formula through sintering simulation and process adaptation, and feeds back optimization. According to the invention, the full-process intelligentization of the silver paste formula from data processing to optimization verification is realized, the conductivity and other performances of the silver paste are accurately improved, the research and development cost is reduced, the period is shortened, the suitability of the formula process is enhanced, and the research and development and industrial upgrading of the silver paste are promoted.

Description

Technical Field

[0001] This invention relates to the field of silver paste preparation technology, specifically to a silver paste conductivity performance data modeling and formulation optimization system based on machine learning. Background Technology

[0002] In the photovoltaic and electronics manufacturing industries, silver paste, as a key electrode material, directly affects the conductivity, reliability, and manufacturing cost of products. In recent years, with the rapid development of high-efficiency solar cell technologies such as TOPCon and HJT, and the evolution of electronic devices towards miniaturization and high integration, the market has placed higher demands on the conductivity, adhesion, and curing speed of silver paste, while simultaneously needing to balance cost control and process adaptability. Traditional silver paste formulation development models face numerous challenges.

[0003] Currently, the optimization of silver paste formulations largely relies on engineers' experience-based trial and error and small-scale experimental adjustments. This approach is not only time-consuming and labor-intensive, with development cycles lasting months or even years, but it also struggles to systematically consider the complex synergistic effects between various components, resulting in low optimization efficiency, high costs, and an inability to meet the rapidly iterating market demands. Furthermore, due to a lack of in-depth analysis of the relationship between process parameters and formulation performance, optimized formulations often exhibit mismatches with actual production processes, such as excessively high sintering temperatures leading to cell damage, or unsuitable viscosity affecting coating uniformity.

[0004] As machine learning and artificial intelligence technologies are increasingly applied in materials research and development, some studies have attempted to introduce data-driven methods into silver paste formulation optimization. However, most existing technologies focus only on the impact of a single component on performance, neglecting the interactions of multiple components and the coupling effects of process parameters; or they fail to effectively incorporate actual production constraints during the optimization process, resulting in formulations lacking engineering feasibility. Furthermore, traditional methods often lack closed-loop verification mechanisms, making it impossible to feed experimental results back into the research and development process, hindering continuous optimization.

[0005] Therefore, a machine learning-based data modeling and formulation optimization system for the conductivity properties of silver paste is proposed to address the above problems. Summary of the Invention

[0006] The purpose of this invention is to provide a machine learning-based system for modeling the conductivity of silver paste and optimizing its formulation, in order to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A machine learning-based data modeling and formulation optimization system for silver paste conductivity includes: Data acquisition and storage module: used to acquire the original dataset of silver paste formulation, including historical formulation components, process parameters and performance index data. The performance index data includes at least conductivity, adhesion, curing time and photoelectric conversion efficiency. Data preprocessing module: Normalizes the original dataset and generates high-dimensional feature vectors based on feature engineering; Feature Influence Analysis Module: Uses machine learning models to calculate the feature influence coefficients of each component on the target performance, and extracts high-influence and low-influence components based on SHAP values; Formulation optimization module: Combining the percentage content constraint of high-impact components and the mean filling strategy of low-impact components, a multi-objective optimization algorithm is used to generate a formulation scheme that meets the preset performance. Simulation verification module: Predicts the electrical performance output of the optimized formula through a sintering process simulation model and feeds it back to the formula optimization module for iterative adjustment.

[0008] As a preferred approach, the operations performed by the feature impact analysis module include: Calculate the characteristic influence coefficient of each component on the target performance: For each component i, iterate through all samples j, multiply the characteristic value of the component by the difference between the performance value of the corresponding sample and the benchmark coefficient, and sum the calculation results of all samples. Quantifying the synergistic effect between components: For any two components i and j, iterate through all feature subsets S that do not contain i and j, calculate the change in the predicted value of the machine learning model when i and j are added to subset S, subtract the change in the predicted value of the model when i is added to subset S, subtract the change in the predicted value of the model when j is added to subset S, and finally add the predicted value of the model on subset S; multiply the calculation result by the weight coefficient and sum over all subsets. The weight coefficient is determined by the factorial of the subset size, the total number of features minus the factorial of the subset size minus two, and the factorial of the total number of features minus one. Optimize feature weights: Multiply the original feature influence coefficient of the component by the exponential decay factor, which is based on the natural constant e and has an exponent of negative decay intensity coefficient multiplied by the content ranking value of the component in the formulation, and add the sum of the interaction influence values ​​of the component with all other components. Components whose final feature influence value is greater than a preset threshold are identified as high-influence components, while the rest are identified as low-influence components.

[0009] As a preferred option, the operation of the formula optimization module includes: Define a sequence of percentage content levels for high-impact components, and generate a full combination formulation space based on the Cartesian product; Perform multi-objective optimization: Maximize the predicted conductivity value; Minimize the total content of nano-silver powder; Satisfy dynamic process constraints: For each process parameter k, calculate the absolute difference between the current formula parameter value and the threshold, subtract the adaptive tolerance coefficient, and take the positive part, which is required to be no more than zero; Dynamically adjust the tolerance coefficient: multiply the initial tolerance coefficient by the adjustment factor raised to the power of t. The adjustment factor is 1 plus the learning rate multiplied by the proportion of individuals violating the constraints to the population size. The process constraints include a viscosity range of 70–130 Pa·s, a fineness index of ≤14 micrometers, and a sintering temperature of 100–200℃.

[0010] As a preferred option, the simulation verification module includes: Sintering process simulation unit: Based on finite element analysis, simulate the silver-copper alloying process of silver paste in a sintering environment of 100-200℃, and predict the interfacial contact resistance and anti-migration performance. Laser-enhanced contact optimization technology adaptation unit: The mass ratio of etherified melamine-formaldehyde resin to polyether polyol resin in the optimized formula is 25-45%:30-55% to adapt to the laser-enhanced contact optimization process of TOPCon cells.

[0011] As a preferred approach, the data preprocessing module performs the following operations: Imputing missing data: Use the average of the predictions of all decision trees in the random forest for the known feature values ​​as the missing feature values; Detect and remove outliers identified by the isolated forest algorithm; Enhanced characteristics include the mass ratio of nano-silver powder to photovoltaic glass powder, and the product of sintering temperature and curing time.

[0012] As a preferred approach, the machine learning model employs a two-layer ensemble structure: Primary prediction layer: XGBoost regression model predicts conductivity; Random forest classification models predict weldability; Meta-model layer: Using the output of the primary prediction layer as features, a comprehensive performance score is generated through a support vector machine.

[0013] As a preferred option, dynamic process constraints include: Viscosity range: 70–130 Pa·s; Fineness specification: ≤14 micrometers; Sintering temperature range: 100–200℃.

[0014] As a preferred option, the silver paste formulation includes the following components and their content constraints: Nano silver powder: 70–80 parts, D50 particle size 0.5–4 micrometers; Nano copper powder: 0.1–1 part, D50 particle size 30–100 nanometers; Photovoltaic glass powder: 1–5 parts, comprising 50–60% by weight lead oxide, 15–25% by weight silicon dioxide, and 10–20% by weight boron trioxide; Organic carrier: 15–30 parts, comprising etherified melamine-formaldehyde resin and polyether polyol resin.

[0015] As a preferred option, the system's silver paste formulation optimization method includes: Input target performance constraint: conductivity ≥ 5 × 10 5 Siemens / meter, welding tensile strength ≥ 3 Newtons / 1.6 mm; Identify high-impact components and narrow down their content candidate set; A candidate formula set is generated, and the Pareto optimal solution set is screened through sintering simulation. The experiment validates the optimized formula and feeds back to the data acquisition module.

[0016] As can be seen from the technical solutions provided by the present invention above, the beneficial effects of the machine learning-based silver paste conductivity data modeling and formulation optimization system provided by the present invention are: Precisely optimize the formula to improve product performance: Based on historical data and machine learning algorithms, we conduct in-depth analysis of the impact of each component on the performance indicators of silver paste, such as conductivity and adhesion. We not only consider the role of individual components, but also quantify the synergistic effect between components. The candidate formulas generated by the formula optimization module can be screened by the simulation verification module to achieve precise improvement in the conductivity, welding tensile strength and other properties of silver paste, meeting the stringent requirements of photovoltaic, electronics and other fields for high-performance silver paste. Data-driven decision-making reduces R&D costs: Abandoning the traditional experience-based trial-and-error R&D model, the system utilizes a large amount of historical data accumulated by the data acquisition and storage module, combined with data preprocessing module for data cleaning, transformation, and feature enhancement, to provide high-quality data for subsequent analysis; on this basis, the feature impact analysis module identifies key influencing factors, the formula optimization module efficiently searches for the optimal formula, and the simulation verification module reduces unnecessary physical experiments, significantly shortening the R&D cycle and reducing raw material waste and experimental costs. Dynamically adapting to needs and enhancing process adaptability: The formulation optimization module adopts a multi-objective optimization algorithm, which can simultaneously consider multiple objectives such as maximizing conductivity and minimizing the content of nano-silver powder, and dynamically adjust the target weights according to actual needs; at the same time, combined with process parameter constraints and adaptive tolerance mechanisms, it ensures that the optimized formulation meets the process requirements such as viscosity, fineness, and sintering temperature in the production process; the simulation verification module optimizes the resin ratio in the formulation for specific application scenarios such as TOPCon batteries, making it suitable for advanced processes such as laser-enhanced contact, and enhancing the applicability of the product under different process conditions; Forming an optimization closed loop to drive continuous improvement: From data collection, feature analysis, and formula optimization to simulation verification and experimental feedback, a complete R&D optimization closed loop is formed; the experimental verification results are fed back to the data collection module to continuously update and enrich the dataset, providing more sufficient data support for subsequent R&D, enabling the system to continuously optimize and adapt to changes in technology development and market demand; Accumulated knowledge and experience facilitate technology transfer: The large amount of data and optimization cases accumulated during system operation have formed a valuable knowledge and experience base; this knowledge can be used to guide the subsequent research and development of silver paste formulations, accelerate the research and development process of new technologies, and also facilitate the transfer of technical knowledge within the enterprise, thereby improving the overall research and development capabilities. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall structure of the machine learning-based silver paste conductivity performance data modeling and formulation optimization system of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific embodiments.

[0020] like Figure 1 As shown, this embodiment of the invention provides a machine learning-based data modeling and formulation optimization system for the conductivity properties of silver paste, including: Data acquisition and storage module: used to acquire the original dataset of silver paste formulation, including historical formulation components, process parameters and performance index data. The performance index data includes at least conductivity, adhesion, curing time and photoelectric conversion efficiency. Data preprocessing module: Normalizes the original dataset and generates high-dimensional feature vectors based on feature engineering; Feature Influence Analysis Module: Uses machine learning models to calculate the feature influence coefficients of each component on the target performance, and extracts high-influence and low-influence components based on SHAP values; Formulation optimization module: Combining the percentage content constraint of high-impact components and the mean filling strategy of low-impact components, a multi-objective optimization algorithm is used to generate a formulation scheme that meets the preset performance. Simulation verification module: Predicts the electrical performance output of the optimized formula through a sintering process simulation model and feeds it back to the formula optimization module for iterative adjustment.

[0021] In this embodiment, the data acquisition and storage module is the "cornerstone" of the machine learning-based silver paste conductivity data modeling and formulation optimization system. Through systematic data acquisition and management, it provides accurate and comprehensive data support for subsequent modeling analysis and formulation optimization. The following will elaborate on this module in terms of its overall functionality, sub-module composition, and key technologies: I. Overall Function Overview: The data acquisition and storage module is responsible for collecting, cleaning, and storing data throughout the entire lifecycle of silver paste formulations. It acquires historical formulation components, process parameters, and performance index data from multiple stages such as production and R&D, and stores them in the database after standardization. At the same time, the module has data verification and update functions to ensure the accuracy and timeliness of the data, laying a solid foundation for the stable operation of the system and the accurate training of the model. II. Submodule Composition and Functions: (a) Multi-source data acquisition unit: Data source integration: Real-time connection with production equipment, laboratory testing instruments and R&D record system to collect multi-dimensional data of silver paste formulation; including the content and particle size parameters of formulation components such as nano silver powder and copper powder; process parameters such as viscosity, fineness and sintering temperature; performance index data such as conductivity, adhesion and photoelectric conversion efficiency, forming a complete dataset; Data interface adaptation: For communication protocols of different devices and systems, industrial standard protocols such as Modbus and OPCUA are adopted, as well as general interfaces such as RESTful API, to achieve seamless data access and transmission; for example, real-time temperature data can be obtained from sintering equipment through the Modbus protocol, and conductivity test results can be retrieved from the detection system through the API interface. (II) Data Preprocessing and Verification Unit: Data cleaning: Clean the collected raw data to remove duplicate records and invalid data; use outlier detection algorithms (such as the 3σ principle) to identify and correct erroneous data to ensure data quality; for example, if the conductivity data of a batch of silver paste exceeds the normal range by 3 times the standard deviation, it is marked as an outlier and corrected or deleted. Format standardization: Convert data from different sources and in different formats into a standard format that the system can recognize; for example, standardize the content data of formulation components to "parts" and the particle size parameter to micrometers (μm), which will facilitate subsequent data processing and analysis. Data validation: Establish a data validation rule base to perform logical validation on key data; for example, check whether the content of nano silver powder is within the set range of 70-80 parts and whether the sintering temperature is within the range of 100-200℃. If the rules are not met, trigger an alarm and prompt data correction. (III) Data Storage and Management Unit: Database architecture design: A hybrid storage architecture combining relational databases (such as MySQL) and non-relational databases (such as MongoDB) is adopted; the relational database is used to store structured formula data, process parameters and performance indicators, which facilitates complex queries and analysis; the non-relational database is used to store unstructured documents, experimental reports, etc., to meet diverse data storage needs; Data Indexing and Retrieval: Indexes are created for commonly used query fields, such as formula ID, batch number, and testing time, to improve data retrieval efficiency; users can quickly query the data they need by keywords, time range, performance index intervals, and other conditions, and fuzzy and combined queries are supported; Data Backup and Recovery: Establish a regular backup strategy, using a combination of full and incremental backups to back up data to local storage devices and the cloud; at the same time, establish a data recovery mechanism to quickly recover data in the event of loss or damage, ensuring data security and integrity; III. Key Technology Principles: (I) Principles of Multi-Source Data Acquisition Technology: Based on Internet of Things (IoT) technology and industrial communication protocols, real-time acquisition of multi-source data from production equipment, testing instruments, etc. is achieved; through device drivers and data acquisition gateways, heterogeneous data generated by different devices are converted into a unified format to ensure data compatibility and transmissibility; using edge computing technology, the raw data is preliminarily processed and filtered at the data acquisition end to reduce the amount of data transmission and improve data acquisition efficiency. (II) Principles of Data Preprocessing Technology: Statistical methods and machine learning algorithms are used for data cleaning and outlier detection; the 3σ principle, based on the principle of normal distribution, determines whether the data belongs to the normal range by calculating the mean and standard deviation of the data; the isolated forest algorithm uses a tree model to divide the data and quickly identify abnormal data; format standardization unifies data of different formats into the system standard format through data mapping and conversion rules to ensure data consistency. (III) Principles of Data Storage Management Technology: Hybrid storage architecture combines the structured query advantages of relational databases with the flexible storage characteristics of non-relational databases; relational databases achieve efficient data querying and transaction processing through table structures and indexing mechanisms; non-relational databases use document storage or key-value pair storage methods, enabling fast storage and retrieval of unstructured data; data backup and recovery technology is based on the principles of data redundancy and version control, ensuring that data can be restored to a correct state in the event of loss or corruption by regularly backing up and storing multiple versions of data. IV. Module Workflow: (a) Initialization phase: After the data acquisition and storage module starts, it loads the data acquisition configuration file, which includes information such as device connection parameters, data acquisition frequency, and interface protocol. Establish communication connections with data sources such as production equipment and testing systems, conduct connection tests, and ensure normal communication. (II) Data Collection Phase: According to the set collection frequency, relevant data on the silver paste formula are obtained in real time from various data sources, including formula components, process parameters and performance indicators. The collected data is initially analyzed and converted to meet the requirements of data preprocessing. (III) Data Preprocessing Stage: The collected data is cleaned to remove duplicate, invalid, and abnormal data. Standardize data formats and unify data units and encoding methods; Validate key data to check whether it conforms to preset rules and ranges, and correct or mark any problems found. (iv) Data storage stage: Based on the data type and structure, the preprocessed data is stored in the corresponding database; structured data is stored in relational databases, and unstructured data is stored in non-relational databases. Index the stored data, optimize the data storage structure, and improve data storage and query efficiency; (v) Data Management Phase: Regularly back up the database, storing the data on local storage devices and in the cloud; Monitor the database's operational status, check metrics such as data storage capacity and query performance, and promptly address database failures and performance issues; Responding to user data query requests, the system retrieves and filters data from the database and returns data results that meet the specified criteria. (vi) Conclusion: When the system stops running or receives a stop command, the data acquisition and storage module stops data acquisition and storage operations, closes the communication connection with the data source, releases system resources, and saves operation records and configuration parameters.

[0022] In this embodiment, the operations performed by the data preprocessing module include: Imputing missing data: Use the average of the predictions of all decision trees in the random forest for the known feature values ​​as the missing feature values; Detect and remove outliers identified by the isolated forest algorithm; Enhanced characteristics include the mass ratio of nano-silver powder to photovoltaic glass powder, and the product of sintering temperature and curing time; Furthermore, the data preprocessing module acts as the "data quality inspector" of the silver paste conductivity data modeling and formulation optimization system. Through in-depth processing and optimization of the raw data, it eliminates data noise and uncovers potential features, laying a solid foundation for the accurate training of subsequent machine learning models and formulation optimization. The following will elaborate on this in terms of functional positioning, core technologies, and workflow: I. Overall Function Overview: The data preprocessing module is mainly responsible for cleaning, transforming, and enhancing the features of the original dataset. It uses various algorithms to normalize the data, addressing issues such as missing values, outliers, and dimensional differences. Simultaneously, based on feature engineering techniques, it derives more representative high-dimensional feature vectors from the original data, improving the data's interpretability of silver paste performance and ensuring the data quality and usability of the input model. II. Submodule Composition and Functions: (a) Data Cleaning Unit: Missing value handling: A random forest imputation algorithm is used, with the following formula: (in, These are missing feature values; The number of decision trees represents the complexity of the model. For the first The prediction function for each tree is based on the observed feature values. The algorithm makes predictions; it constructs multiple decision trees, integrates the prediction results of each tree, and effectively utilizes the correlation between features to fill in missing values, such as accurately estimating the missing data on the content of nano-silver powder. Outlier Detection and Handling: The system employs the Isolation Forest algorithm to isolate data points by constructing a binary tree, quickly identifying data points that deviate from the normal distribution. For detected outliers, such as extreme deviations in the curing time of a batch of silver paste, the system will automatically mark and delete them to prevent them from interfering with subsequent analysis. (ii) Data standardization unit: Normalization: The original data is normalized to convert features with different dimensions (such as viscosity in Pa·s and conductivity in Siemens / meter) to a uniform range (usually [0,1] or [-1,1]); a minimum-maximum normalization formula is used. (in, These are the original eigenvalues. and These are the minimum and maximum values ​​of the feature, respectively. (These are the normalized feature values) to ensure that each feature has equal weight during model training, and to avoid the model being biased towards certain features due to differences in units; Standardization transformation: For features that conform to a normal distribution, use... -score standardization method (in, The original value, The characteristic mean, Standard deviation, (These are standardized values), so that the data mean is 0 and the standard deviation is 1, which improves the convergence speed and stability of the model. (III) Feature Engineering Units: Interaction generation: By calculating the interaction relationships between components, hidden synergistic effects can be discovered; for example, the mass ratio of nano-silver powder to photovoltaic glass powder can be generated, and changes in this ratio may significantly affect the sintering density and conductivity of silver paste; the ratio of etherified melamine-formaldehyde resin to polyether polyol resin in the organic carrier will affect the rheological properties and adhesion of silver paste. Coupling term construction: Based on the correlation between process parameters, process parameter coupling terms are constructed, such as the product of sintering temperature and curing time. This coupling term can reflect the comprehensive influence of different process condition combinations on the performance of silver paste. For example, high temperature short time sintering and low temperature long time sintering may produce different microstructures and performance characteristics. High-dimensional feature generation: Combining dimensionality reduction techniques such as principal component analysis (PCA) and singular value decomposition (SVD), the original features are mapped to a high-dimensional space while retaining the main information of the data, generating more discriminative feature vectors and providing higher-quality input for machine learning models; III. Key Technology Principles: (a) Principles of missing value imputation: The random forest imputation algorithm is based on the idea of ​​ensemble learning, which uses multiple decision trees to fit and predict data. Each decision tree builds a model based on existing complete feature data, and then predicts the missing value samples. Finally, the estimated value of the missing value is obtained by averaging the prediction results. This method can make full use of the nonlinear relationship between data features and has higher accuracy than the traditional mean / median imputation method. (II) Outlier Detection Principle: The Isolation Forest algorithm is based on the assumption that "outliers are easier to isolate in the data space". It constructs a binary tree by randomly selecting features and split points to divide data points into different nodes. Since the number of outliers is small and their distribution is discrete, their paths in the tree are usually short. By calculating the average path length of all data points and setting a threshold, outliers are identified. It has the advantages of high detection efficiency and no need to pre-set distribution assumptions. (III) Characteristic Engineering Principles: The construction of interaction and coupling terms is based on the physicochemical mechanism of silver paste formulation, and the synergistic effect between components or process parameters is revealed through mathematical operations; high-dimensional feature generation utilizes the dimensionality reduction technique in linear algebra to map the original data to a new feature space, thereby reducing the data dimensionality and enhancing the separability of the data and the expressive power of the model. For example, PCA retains the feature information with the largest variance by finding the principal component direction of the data. IV. Module Workflow: (a) Data Reception Stage: The raw dataset is obtained from the data acquisition and storage module, including the composition, process parameters and performance index data of the silver paste formula. The data integrity and format standardization are checked. If there are format errors, the data cleaning unit is triggered to make preliminary corrections. (II) Data Cleaning Stage: The missing value handling algorithm is invoked to fill in the missing values ​​in the original data and generate a complete dataset; The data is scanned using an outlier detection algorithm, outlier data points are marked and deleted, and the cleaned dataset is output. (III) Data Standardization Stage: Based on the data type and distribution characteristics, select normalization or standardization methods to transform the cleaned data and unify the data units and distribution format; Perform a quality check on the standardized data to ensure that the data is within a reasonable range. If any anomalies are found, readjust the standardization parameters. (iv) Feature Engineering Stage: Based on knowledge of silver paste formulation and process experience, pre-defined interaction terms and coupling terms are calculated to expand the original feature set; The expanded feature set is processed using a dimensionality reduction algorithm to generate a high-dimensional feature vector, which serves as the input data for subsequent machine learning models. (v) Output stage: The processed dataset is transferred to the feature impact analysis module, while intermediate processing data is saved for subsequent traceability and model optimization. V. Application Value of the Module: (a) Improve model performance: By cleaning and standardizing the data, noise and dimensional interference are eliminated, avoiding overfitting or underfitting problems in model training; the high-dimensional feature vectors generated by feature engineering significantly enhance the data's ability to interpret the performance of silver paste, enabling machine learning models to more accurately capture the complex relationship between formulation and performance, and improving prediction accuracy and optimization effect. (ii) Reduce computing costs: Standardization reduces the dispersion of data and accelerates the convergence speed of gradient descent during model training; dimensionality reduction technology reduces the dimensionality of data, reduces the computational load and memory usage of the model, and improves the system's operating efficiency while ensuring model performance. (iii) Enhancing the reliability of analysis: A rigorous data cleaning and verification mechanism ensures that the data input into the model is authentic and reliable, avoiding erroneous analysis results caused by abnormal data. For example, accurately filling in the missing values ​​of nano-silver powder content can prevent the model from incorrectly identifying its related factors as key influencing factors, thereby improving the credibility of the formulation optimization conclusions. (iv) Supporting technological innovation: The interaction and coupling terms mined by feature engineering reveal the potential patterns between silver paste formulation and process parameters, providing researchers with new optimization directions. For example, by analyzing the influence of the ratio of nano-silver powder to resin on conductivity, the formulation can be adjusted in a targeted manner, promoting innovation and improvement in silver paste preparation processes.

[0023] In this embodiment, the operations performed by the feature influence analysis module include: Calculate the characteristic influence coefficient of each component on the target performance: For each component i, iterate through all samples j, multiply the characteristic value of the component by the difference between the performance value of the corresponding sample and the benchmark coefficient, and sum the calculation results of all samples. Quantifying the synergistic effect between components: For any two components i and j, iterate through all feature subsets S that do not contain i and j, calculate the change in the predicted value of the machine learning model when i and j are added to subset S, subtract the change in the predicted value of the model when i is added to subset S, subtract the change in the predicted value of the model when j is added to subset S, and finally add the predicted value of the model on subset S; multiply the calculation result by the weight coefficient and sum over all subsets. The weight coefficient is determined by the factorial of the subset size, the total number of features minus the factorial of the subset size minus two, and the factorial of the total number of features minus one. Optimize feature weights: Multiply the original feature influence coefficient of the component by the exponential decay factor, which is based on the natural constant e and has an exponent of negative decay intensity coefficient multiplied by the content ranking value of the component in the formulation, and add the sum of the interaction influence values ​​of the component with all other components. Components whose final feature influence value is greater than a preset threshold are identified as high-influence components, and the rest are identified as low-influence components. Furthermore, the feature impact analysis module serves as the "decision-making brain" of the silver paste conductivity performance data modeling and formulation optimization system. By quantifying the impact of each formulation component and process parameter on the target performance, it provides a scientific basis for formulation optimization. This module integrates traditional statistical methods with advanced machine learning techniques to accurately identify key influencing factors, helping R&D personnel focus on core variables and improve optimization efficiency. The following is a detailed explanation from the perspectives of module architecture, core algorithms, and workflow: I. Overall Function Overview: The feature impact analysis module constructs a multi-dimensional evaluation system to calculate the impact coefficients of each component and process parameter on the performance indicators of silver paste, such as conductivity and adhesion. Based on the SHAP value theory, it quantifies the single-factor effect and synergistic effect, and introduces a content decay mechanism to correct the weights. Finally, it extracts the high-impact components as the core variables for formulation optimization. The module output results directly guide the parameter search space of the formulation optimization module, significantly improving the optimization efficiency and accuracy. II. Submodule Composition and Functions: (I) Basic Influence Coefficient Calculation Unit: Feature contribution metric: based on formula (in, Let be the characteristic influence coefficient of the i-th component on the target performance; Let j be the performance value of the j-th sample; The baseline coefficient is the average performance of the training set. (where N is the number of samples; i is the feature of the i-th component of the j-th sample). Physical meaning: This formula quantifies the linear contribution of a single component change to performance by calculating the deviation of the sample performance from the benchmark value; for example, when the content of nano-silver powder increases, if... If the value is positive and large, it indicates that it has a significant positive impact on improving conductivity. Dynamic baseline adjustment: Automatically updates baseline coefficients based on data from different batches or process conditions. To ensure the accuracy and timeliness of the impact coefficient calculation; (II) Interaction Effect Analysis Unit: Synergy Quantification: An Improved Interactive SHAP Algorithm (in, Components Interaction values ​​with other components; For the set of features of all components; A subset of features; For machine learning model prediction functions; For feature subset The number of elements; (The number of elements in the feature set M of all components). Application scenarios: For example, when nano-silver powder and photovoltaic glass powder work together, this formula can be used to calculate the synergistic effect of the change in their ratio on the formation of the silver particle network after sintering, revealing hidden relationships that traditional single-factor analysis cannot discover; Multi-order interactive calculation: Supports calculation of third-order and above interactive effects, such as the complex interaction between nano silver powder, copper powder and organic carrier, and comprehensively captures the nonlinear relationship in the formulation system; (III) Weight Optimization Unit: Content decay correction: Introducing a component effect decay factor (in, Let be the final characteristic influence coefficient of the i-th component; Let i be the characteristic influence coefficient of the i-th component on the target performance; This is the attenuation intensity coefficient, with a default value of 0.2; Rank the content of component i in the formula, with 1 being the highest content; (This refers to the interaction value between component i and component j). Practical significance: For low-content components (such as nano copper powder accounting for only 0.1-1 parts), the weight of its basic influence coefficient is reduced by the attenuation factor, which avoids overfitting caused by small content fluctuations and improves the generalization ability of the model. Adaptive parameter adjustment: Dynamically adjusts the attenuation intensity coefficient according to the characteristics of different formulation systems. For example, appropriately increasing the sensitivity of a highly sensitive system Value, to enhance the attenuation effect; (iv) Component classification and screening unit: Threshold division mechanism: setting an influence coefficient threshold ,Will The components are classified as high-impact components, and the rest as low-impact components; Application Case: In typical silver paste formulations, nano-silver powder and organic carriers are usually identified as high-impact components, while some additives may be low-impact components. Dynamic threshold update: The threshold is updated periodically based on historical optimization data and domain knowledge. This ensures the accuracy and adaptability of the classification results; III. Key Technology Principles: (I) Application of SHAP value theory: The SHAP (SHapley Additive ex Planations) value originates from cooperative game theory. It provides a theoretical basis for feature influence analysis by calculating the contribution of each feature to the model prediction. The improved SHAP algorithm, based on the traditional method, introduces the idea of ​​permutation and combination from combinatorics to efficiently calculate the interaction effect between high-dimensional features, thus solving the problem of high computational complexity of the traditional method. (II) Attenuation factor design principle: Attenuation factor The design is based on the physical principle that "the lower the content of a component, the smaller its impact on the overall performance"; nonlinear decay is achieved through an exponential function, which retains the potential influence of low-content components while avoiding their excessive interference with model judgment; for example, although nano copper powder has an enhancing effect on conductivity, its proportion is extremely low, and its weight can be reasonably reduced through the decay factor. (III) Multi-indicator integration mechanism: The module will include the basic influence coefficient. Interaction impact value Combined with attenuation correction, the final influence coefficient is formed. This fusion mechanism comprehensively considers single-factor effects, synergistic effects, and content distribution. Compared with traditional methods that only focus on main effects, it can more accurately reflect the real physicochemical process. IV. Module Workflow: (a) Data preparation stage: Obtain standardized feature vectors from the data preprocessing module, including formulation components, process parameters, etc. Load the pre-trained machine learning model (such as XGBoost, Random Forest, etc.) for subsequent SHAP value calculation; (II) Basic Impact Calculation Stage: Calculate the deviation of each sample's performance value from the benchmark value. ; Iterate through all samples, according to the formula Calculate the basic influence coefficient of each component ; Construct an influence coefficient matrix to record the degree of influence of each component on different performance indicators; (III) Interaction Effect Analysis Stage: For each component, construct a feature subset. Calculate its interaction values ​​with other components. ; Monte Carlo sampling is used to accelerate the calculation of higher-order interaction effects, reducing computational complexity while maintaining accuracy; Generate an interaction effect heatmap to visually demonstrate the strength and direction of the synergistic effect between the components; (iv) Weight optimization stage: Calculate the attenuation factor based on the component content order in the formulation. ; The final impact coefficient is calculated by combining the basic impact coefficient and the interaction impact value. ; The influence coefficient is normalized to facilitate subsequent threshold division and comparison; (V) Component Classification and Output Stage: Comparison of each component With threshold Divide into high- and low-impact components; Generate an impact analysis report, including the ranking of the impact coefficients of each component, the results of the interaction effect analysis, etc. The list of high-impact components is output to the formulation optimization module as a constraint on the optimization variable space; V. Application Value of the Module: (a) Focusing on core variables: By accurately identifying high-impact components, researchers can reduce their optimization focus from dozens of variables to a few key ones, such as the content of silver nanopowder and the ratio of organic carriers, which significantly improves optimization efficiency and reduces R&D costs. (II) Revealing the hidden mechanisms: Interaction effect analysis revealed the synergistic effect between components, such as the optimal mass ratio of nano-silver powder to photovoltaic glass powder, providing a new theoretical basis for formulation design and promoting the transformation from "experience-based trial and error" to "scientific design". (III) Improving model accuracy: The attenuation factor mechanism effectively solves the problem of overfitting of low-content components, making the model more focused on the truly key influencing factors and improving the accuracy and generalization ability of the performance prediction model. (iv) Guiding process improvement: The high-impact components and process parameters output by the module can provide optimization directions for actual production; for example, if the sintering temperature is found to be a high-impact parameter, the sintering process can be optimized in a targeted manner to improve the stability of product quality.

[0024] In this embodiment, the operation of the formula optimization module includes: Define a sequence of percentage content levels for high-impact components, and generate a full combination formulation space based on the Cartesian product; Perform multi-objective optimization: Maximize the predicted conductivity value; Minimize the total content of nano-silver powder; Satisfy dynamic process constraints: For each process parameter k, calculate the absolute difference between the current formula parameter value and the threshold, subtract the adaptive tolerance coefficient, and take the positive part, which is required to be no more than zero; Dynamically adjust the tolerance coefficient: multiply the initial tolerance coefficient by the adjustment factor raised to the power of t. The adjustment factor is 1 plus the learning rate multiplied by the proportion of individuals violating the constraints to the population size. The process constraints include a viscosity range of 70–130 Pa·s, a fineness index of ≤14 microns, and a sintering temperature of 100–200℃. Furthermore, the formulation optimization module serves as the "innovation engine" of the silver paste conductivity data modeling and formulation optimization system. Through a multi-objective intelligent optimization algorithm, it searches for the optimal formulation combination in a high-dimensional parameter space. Based on the results of feature influence analysis, and combined with process constraints and performance objectives, this module generates a Pareto optimal solution set that meets industrial requirements, significantly improving R&D efficiency and product performance. The following sections elaborate on the module architecture, core algorithms, and workflow: I. Overall Function Overview: The formulation optimization module receives the list of high-impact components output by the feature influence analysis module and constructs a constraint-enhanced multi-objective optimization model. By setting conflicting objectives such as maximizing conductivity and minimizing nano-silver powder content, combined with process parameter constraints, the improved NSGA-II algorithm is used to perform intelligent search in the formulation space. At the same time, an adaptive tolerance mechanism is introduced to dynamically adjust the constraint boundary, balance optimization accuracy and feasibility, and finally output a set of Pareto optimal formulation schemes for R&D personnel to choose from. II. Submodule Composition and Functions: (a) Optimize the target building unit: Multi-objective function design: Objective function 1: Predicted conductivity (maximum) (where, (This is a formula feature vector, containing the content of each component and process parameters) Objective function 2: Minimize the total content of nano silver powder. Physical significance: To achieve a balance between performance and cost by minimizing the amount of precious metal silver used while ensuring conductivity. Target weights are dynamically adjusted: the relative importance of each target is dynamically adjusted according to market demand and cost fluctuations. For example, when the price of silver rises, the weight of the target of minimizing the content of nano silver powder is increased. (ii) Constraint Modeling Unit: Process parameter constraints: (in, For the first A dynamic tolerance function for each process constraint; For the current formula Item process parameter values; For process parameter thresholds; The adaptive tolerance coefficient has an initial value of 5% of the threshold. Typical constraints: viscosity range 70–130 Pa·S, fineness ≤14 μm, sintering temperature 100–200℃; Component content constraints: Nano silver powder: 70–80 parts, D50 particle size 0.5–4 μm; Nano copper powder: 0.1–1 part, D50 particle size 30–100 nm; Photovoltaic glass powder: 1–5 parts, specific component requirements; Organic carrier: 15–30 parts, with specific resin ratio; (III) Optimization Algorithm Execution Unit: Constraint-enhanced NSGA-II algorithm: We employ an elite retention strategy and non-dominated sorting to maintain the Pareto optimal solution set; Adaptive crossover and mutation operators are introduced to design finer search step sizes for high-impact components, such as a search precision of 0.1 parts for the content of nano silver powder; Adaptive tolerance adjustment: (in, For the first The tolerance coefficient for iteration; This is the initial tolerance factor; This is the learning rate, with a default value of 0.05. The number of individuals that violated the constraints; Population size; (Number of iterations); Dynamic adjustment mechanism: In the early stage of iteration, the tolerance is increased to explore a wider space, and in the later stage, the tolerance is tightened to improve the accuracy of the solution; (iv) Results Evaluation and Output Unit: Pareto front analysis: Calculate the distribution density and spacing indices of the solution set to evaluate the diversity and convergence of the solution set; Decision support system: It provides a multi-dimensional visualization interface to show the trade-off between conductivity and silver powder content; The process feasibility scores of each scheme are marked. For example, the scheme with a sintering temperature close to the optimal range of industrial equipment receives a higher score. Experiment Prioritization: Rank the optimization schemes according to technical feasibility and expected benefits, and output a recommended experiment list; III. Key Technology Principles: (I) Multi-objective optimization theory: Pareto optimality theory is used to handle conflicting objectives. A set of solutions that cannot be further improved for any objective without harming other objectives is generated through non-dominated sorting. This method avoids the limitations of traditional single-objective optimization and provides researchers with a flexible decision space. (ii) Adaptive tolerance mechanism: By dynamically adjusting the constraint boundaries, the algorithm balances exploration and development capabilities; the larger tolerance in the early stage allows the algorithm to escape local optima, while the smaller tolerance in the later stage ensures the accuracy of the solution, making it particularly suitable for handling complex nonlinear optimization problems. (III) Elite Retention Strategy: In each generation of evolution, a certain proportion of the best individuals are retained to directly enter the next generation to prevent excellent solutions from being lost during the evolution process and to accelerate the convergence of the algorithm. IV. Module Workflow (a) Initialization phase: Load the list of high-impact components and process constraints to determine the optimization variable space; Set algorithm parameters: population size (default 100), maximum number of iterations (default 200), crossover probability (0.8), mutation probability (0.2), etc.; Initial tolerance coefficient To set initial flexibility space for each process constraint; (ii) Population formation stage: An initial population is randomly generated within the optimization variable space to ensure that each individual satisfies the basic constraints. Perform a feasibility check on the initial population and remove individuals that violate hard constraints (such as component content exceeding the range); (III) Iterative Optimization Phase: Fitness assessment: Calculate the objective function value and constraint violation degree for each individual; Non-dominated ranking: stratifying a population and dividing individuals into different non-dominated ranks; Crowding degree calculation: Calculate the crowding degree distance between individuals within the same level to maintain solution set diversity; Selection operation: The tournament selection method is adopted, which combines non-dominance level and crowding degree to select parent individuals; Crossover and Mutation: Perform simulated binary crossover and polynomial mutation operations on the parent individuals to generate offspring; Constraint handling: Adjust the tolerance coefficient based on the current iteration number. Assess the feasibility of offspring; Elite preservation: Merging parent and offspring generations, retaining the best individuals to form the new generation of the population; Termination criteria: The algorithm terminates when the maximum number of iterations is reached or the solution set converges; (iv) Post-processing stage of results: Cluster analysis was performed on the final Pareto solution set to identify representative formulation schemes; Calculate the robustness index of each scheme and assess its sensitivity to parameter fluctuations; By combining domain knowledge, solutions that are not technically feasible or too costly are eliminated; (v) Output stage: Generate an optimization report, including the Pareto front plot, performance predictions for each scheme, and grouping ratios; Provide decision-making suggestions, and indicate recommended experimental schemes and their expected benefits; The optimization results are fed back to the simulation verification module for further verification. V. Application Value of the Module: (a) Breaking through the limitations of traditional trial and error: By using intelligent algorithms to quickly search in high-dimensional space, the R&D cycle is significantly reduced, shortening the traditional optimization period from several months to several weeks or even several days. (ii) Achieving dual optimization of performance and cost: While maintaining conductivity, the amount of nano-silver powder used can be reduced, for example, by decreasing the silver content from 78% to 75%, while keeping the conductivity ≥5×10⁻⁶. 5 S / m significantly reduces raw material costs; (III) Improving process feasibility: The adaptive tolerance mechanism ensures that the optimized solution conforms to the actual industrial production. For example, the viscosity of the generated formula automatically meets the process requirements of 70–130 Pa·S, reducing the cost of subsequent adjustments. (iv) Support innovative formulation design: By revealing the complex interactions between components, we discovered formulation combinations that are difficult to conceive of through traditional experience. For example, a specific ratio of nano-silver powder to photovoltaic glass powder can significantly improve high-temperature stability.

[0025] In this embodiment, the simulation verification module includes: Sintering process simulation unit: Based on finite element analysis, simulate the silver-copper alloying process of silver paste in a sintering environment of 100-200℃, and predict the interfacial contact resistance and anti-migration performance. Laser-enhanced contact optimization technology adaptation unit: Optimize the mass ratio of etherified melamine-formaldehyde resin to polyether polyol resin in the formulation to 25-45%:30-55% to adapt to the laser-enhanced contact optimization process of TOPCon cells; Furthermore, the simulation verification module serves as a "virtual laboratory" for the silver paste conductivity data modeling and formulation optimization system. Through digital simulation and technology adaptation, it conducts performance pre-evaluation and process feasibility verification of the optimized silver paste formulation. This module combines physical simulation with cutting-edge industry technologies to construct a full-chain verification system from microstructure evolution to macroscopic performance output, effectively reducing the number of physical experiments and accelerating the R&D process. The following will elaborate on its functional architecture, core technologies, and workflow: I. Overall Function Overview: The simulation verification module takes the candidate formulations output by the formulation optimization module as input and simulates the physicochemical changes of silver paste during sintering through finite element analysis to predict key indicators such as interfacial contact resistance and anti-migration performance. At the same time, for specific application scenarios such as TOPCon batteries, the module uses laser-enhanced contact optimization technology to adjust the resin ratio in the formulation and verify the applicability of the formulation in actual processes. The module feeds back the simulation results to the formulation optimization module, forming a closed loop of "optimization-verification-iteration" to ensure that the final formulation meets both performance and process requirements. II. Submodule Composition and Functions: (I) Sintering process simulation unit: Multiphysics Modeling: Based on the finite element analysis method, a multiphysics coupled model of the silver paste sintering process is constructed, including: Heat conduction model: Simulates the temperature distribution inside the silver paste under sintering conditions of 100–200℃, using Fourier's law of thermal conductivity. (in, For gradient operators, Thermal conductivity, For temperature, As an internal heat source, For density, For isobaric specific heat capacity, (Time) The effect of temperature on the silver-copper alloying reaction rate was analyzed; Mass diffusion model: based on Fick's second law (in, For substance concentration, (where is the diffusion coefficient) to simulate the diffusion behavior of silver and copper atoms at high temperatures and predict the degree of alloying and microstructure evolution. Stress-strain model: Considering the difference in the thermal expansion coefficient of the materials, calculate the thermal stress generated during sintering and evaluate the bonding stability between the silver paste and the substrate; Performance prediction: Interfacial contact resistance prediction: Based on the microstructure of the sintered silver paste, the contact resistance between the electrode and the silicon wafer is calculated using the transmission line model (TLM) to evaluate the conductivity. Anti-migration performance analysis: By simulating the migration path of silver ions under the influence of electric field and temperature, the reliability of silver paste in long-term use is predicted to avoid the risk of electrode short circuit; Visualization Analysis: Generates dynamic simulation videos of the sintering process, intuitively displaying the changes in the internal structure of the silver paste; outputs visualized data such as temperature cloud maps and concentration distribution to assist R&D personnel in analyzing factors affecting formula performance; (II) Laser-enhanced contact optimization technology adaptation unit: Process requirements analysis: For the laser-enhanced contact process of TOPCon cells, the special requirements for silver paste are clarified, such as high laser absorption rate and rapid curing characteristics; Resin formulation optimization: Parameter scanning: Within the ratio range of etherified melamine formaldehyde resin (25–45%) and polyether polyol resin (30–55%), the laser absorption efficiency, coefficient of thermal expansion and other parameters of silver paste under different ratios were simulated. Response surface methodology: A mathematical model of the compatibility between resin ratio and laser process is constructed using response surface methodology to analyze the interaction of various factors and find the optimal ratio combination. For example, when the proportion of etherified melamine formaldehyde resin is 35% and the proportion of polyether polyol resin is 45%, the silver paste has the highest absorption rate of laser energy and can quickly form a dense conductive layer after laser irradiation. Process window verification: Simulate the silver paste reaction under different process parameters such as laser power (50–200W) and scanning speed (100–500mm / s) to determine the process parameter window and ensure the stability of the formula in actual production; (III) Verification Result Feedback Unit: Performance evaluation: Compare the simulated predicted performance indicators such as conductivity and contact resistance with the target requirements to determine whether the formula meets the standards; if it does not meet the requirements, generate a performance deviation report and mark the key influencing factors. Iterative feedback: The verification results are fed back to the formulation optimization module to trigger a new round of optimization; for example, if the simulation shows that the anti-migration performance of a certain formulation is insufficient, the content of nano copper powder or glass powder composition is adjusted and a new candidate formulation is generated. Data archiving: Simulation process data and verification results are stored in a database to form a historical case library, providing a reference for subsequent research and development; III. Key Technology Principles: (a) Finite element analysis technique: Based on the variational principle, the continuous physical field problem is discretized into a finite number of elements for solution; by dividing the grid, the silver paste model is decomposed into a large number of tiny elements, and the physical equations are solved approximately in each element. Finally, the overall physical field distribution is obtained by integration, realizing high-precision simulation of the sintering process. (ii) Transmission Line Model (TLM): By fabricating silver electrodes with different spacings on the surface of a silicon wafer and measuring the resistance between the electrodes, the interface contact resistance and the thin film resistance are inversely derived using a mathematical model. This model assumes that the current is uniformly distributed at the interface between the electrode and the silicon wafer. By linearly fitting the experimental data, the contact resistance parameters are extracted, which is a common method for evaluating electrode contact performance. (III) Response Surface Methodology: Based on fitting the functional relationship between the response variable and multiple factors using a multiple quadratic regression equation, this method obtains data points by designing experiments and constructs a response surface model. It can intuitively demonstrate the interaction between factors, quickly locate the optimal parameter combination, and is widely used in multi-parameter optimization scenarios. IV. Module Workflow: (a) Data Reception Stage: Candidate silver paste formulations are obtained from the formulation optimization module, including component content (nano silver powder, resin, etc.) and process parameters (sintering temperature, laser power, etc.). (II) Sintering process simulation stage: Model building: Based on the formula parameters, a finite element model of the silver paste sintering process is established, and material properties and boundary conditions (temperature, pressure, etc.) are set. Simulation calculation: Solve multiphysics coupling equations to simulate temperature changes, material diffusion, stress distribution, etc. during the sintering process; Performance prediction: Based on simulation results, calculate indicators such as interface contact resistance and anti-migration performance; (III) Laser process adaptation stage: Parameter settings: Determine the range of laser process parameters and the initial value of resin ratio; Formulation optimization: Find the optimal resin ratio through parameter scanning and response surface modeling; Process validation: Simulate the performance of silver paste under different laser process conditions to determine the process window; (iv) Results Evaluation Phase Performance comparison: Compare the simulation results with the target performance indicators to determine the feasibility of the formulation; Problem diagnosis: If the formula does not meet the standards, analyze the simulation data to pinpoint the performance bottleneck (such as excessively high contact resistance, incomplete resin curing, etc.). (v) Feedback and Iteration Phase: Results feedback: Send the verification results and problem analysis report to the formula optimization module; Optimization and iteration: The formula optimization module adjusts the formula based on feedback, regenerates candidate solutions, and returns to the sintering process simulation stage until the performance requirements are met; (vi) Output stage: Output a detailed simulation report of the qualified formula, including performance data, process parameter suggestions, microstructure analysis, etc., to provide a basis for subsequent experimental verification; V. Application Value of the Module: (a) Reduce R&D costs: By replacing some physical experiments with virtual simulation, the consumption of raw materials and the time spent on equipment can be reduced, which is expected to reduce R&D costs by 30%-50%. (ii) Accelerate product iteration: Rapidly verify the feasibility of a formula, shorten the R&D cycle, reduce the time from concept to mass production of new products by 20%-40%, and enhance the company's market competitiveness; (III) Ensuring process reliability: Specialized adaptations are made for advanced processes such as TOPCon batteries to ensure that the silver paste formula is seamlessly integrated with the actual production process and to avoid product quality problems caused by process mismatch. (iv) Knowledge accumulation and reuse: Establish a simulation case library to digitally record successful and unsuccessful experiences, provide reference for subsequent research and development, and accelerate knowledge transfer and innovation.

[0026] In this embodiment, the system's silver paste formulation optimization method includes: Input target performance constraint: conductivity ≥ 5 × 10 5 Siemens / meter, welding tensile strength ≥ 3 Newtons / 1.6 mm; Identify high-impact components and narrow down their content candidate set; A candidate formula set is generated, and the Pareto optimal solution set is screened through sintering simulation. Experiments were conducted to verify and optimize the formula, and the results were fed back to the data acquisition module. Furthermore, the silver paste formulation optimization method, centered on data-driven and model-based calculations, transforms target performance requirements into practically usable silver paste formulations through a systematic process. This method relies on the collaborative operation of various modules within the system to achieve closed-loop management from initial constraint setting to final optimized formulation output, providing a scientific and efficient solution for silver paste R&D and production. The following will elaborate on this method in detail, focusing on its process framework and key steps: I. Overall Process Overview: The silver paste formulation optimization method starts with clear target performance constraints, uses the feature influence analysis module to identify key influencing factors, generates candidate formulations through the formulation optimization module, and then uses the simulation verification module to screen out the solutions that meet the requirements. Finally, the optimization loop is completed through experimental verification. This method comprehensively uses machine learning algorithms, multi-objective optimization strategies and simulation technology to achieve balanced optimization of silver paste formulations in multiple dimensions such as conductivity, cost and process adaptability. II. Detailed Step-by-Step Explanation: (a) Input target performance constraints: Performance index setting: Define the core performance indexes that the silver paste needs to achieve, such as specific numerical requirements for conductivity, welding pull force, etc., and supplement other indicators such as adhesion, curing time, etc. according to actual needs; Process parameter constraints: Determine the limiting conditions of the production process, such as viscosity range, fineness index, sintering temperature range, etc., to ensure that the optimized formula meets the actual production conditions; (ii) Identify high-impact components by calling the feature impact analysis module: Data processing and calculation: The raw dataset containing historical formula components, process parameters and performance indicators is processed by the data preprocessing module and then input into the feature influence analysis module. This module calculates the influence coefficient of each component on the target performance through quantitative analysis, and considers the synergistic effect between components. At the same time, it optimizes and adjusts the influence weight of low content components. High-impact component identification: Based on the calculation results, high-impact components that have a significant impact on the target performance are screened out, and their content candidate set is determined to narrow down the parameter space for subsequent optimization; (iii) Generate a candidate formula set through the formula optimization module: Formulation space construction: Based on the content candidate set of high-impact components, set their percentage content level value sequence to generate a full combination formulation space to ensure coverage of all possible formulation combinations; Multi-objective optimization solution: Intelligent optimization algorithm is used for multi-objective optimization, which simultaneously considers multiple objectives such as maximizing conductivity and minimizing the total content of nano-silver powder; during the optimization process, the elasticity range of the constraints is dynamically adjusted to balance the exploration and convergence in the optimization process, ensuring that the generated formula meets both performance requirements and process feasibility. Candidate recipe generation: Through iterative calculation by the algorithm, a set of Pareto optimal candidate recipes is output, and each recipe achieves a different degree of balance between different objectives; (iv) The Pareto optimal solution set is output after being filtered by the simulation verification module: Sintering process simulation: Using the sintering process simulation unit of the simulation verification module, the physical and chemical changes of silver paste under actual sintering environment are simulated, key indicators such as interfacial contact resistance and anti-migration performance are predicted, and the stability of the formulation in the actual process is evaluated. Laser process adaptation verification: For specific application scenarios, the mass ratio of resin in the formula is optimized through the laser-enhanced contact optimization technology adaptation unit to verify the adaptability of the formula to the laser-enhanced contact process; Results screening and output: Based on the simulation results, the formulations that meet the target performance constraints and process requirements are screened out, and the final Pareto optimal solution set is output to provide decision-making basis for R&D personnel; (v) Conduct experimental verification of the optimized formula and feed the results back to the data acquisition module: Experimental verification: Silver paste samples were prepared according to the optimized formula, and actual performance tests were conducted to verify the accuracy of the simulation results and the feasibility of the formula; Data feedback and iteration: The experimental results are fed back to the data acquisition and storage module to update the original dataset. If the experimental results deviate from the expectations, the system can restart the optimization process, adjust the parameters and algorithms, and carry out a new round of formula optimization to form a closed loop of continuous improvement.

[0027] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A machine learning based silver paste conductive performance data modeling and formulation optimization system, characterized in that: Comprise: Data acquisition and storage module: for obtaining the original data set of silver paste formula, including historical formula components, process parameters and performance index data, the performance index data at least contains conductivity, adhesion, curing time and photoelectric conversion efficiency; Data preprocessing module: normalizing the original data set, and generating high-dimensional feature vector based on feature engineering; Feature influence analysis module: using machine learning model to calculate the feature influence coefficient of each component on the target performance, and extracting high influence components and low influence components based on SHAP value; Formula optimization module: combining the percentage content constraint of high influence components and the mean filling strategy of low influence components, generating formula scheme meeting the preset performance through multi-objective optimization algorithm; Simulation verification module: predicting the electrical performance output of the optimized formula through the sintering process simulation model, and feeding back to the formula optimization module for iterative adjustment.

2. The machine learning based silver paste conductive performance data modeling and formulation optimization system of claim 1, wherein: The operations performed by the feature influence analysis module include: Calculating the feature influence coefficient of each component on the target performance: for each component i, traverse all samples j, multiply the feature value of the component by the difference between the performance value of the corresponding sample and the baseline coefficient, and accumulate the calculation results of all samples to obtain the sum; Quantifying the synergistic effect between components: for any two components i and j, traverse all feature subsets S that do not contain i and j, calculate the change amount of the prediction value of the machine learning model when i and j are added to subset S, subtract the change amount of the prediction value of the model when i is added to subset S, subtract the change amount of the prediction value of the model when j is added to subset S, and finally add the prediction value of the model on subset S; multiply the calculation result by a weight coefficient, and sum all subsets, wherein the weight coefficient is determined by the factorial of the subset size, the factorial of the total feature number minus two times the subset size, divided by the factorial of the total feature number minus one; Optimizing feature weight: multiplying the original feature influence coefficient of the component by an exponential decay factor, which is based on the natural constant e, with the index being the negative decay strength coefficient multiplied by the content ranking value of the component in the formula, plus the sum of the interaction values between the component and all other components; Extracting components with final feature influence values greater than a preset threshold as high influence components, and the rest as low influence components.

3. The machine learning based silver paste conductive performance data modeling and formulation optimization system of claim 1, wherein: The operations of the formula optimization module include: Setting the percentage content level value sequence of high influence components, generating the full combination formula space based on Cartesian product; Performing multi-objective optimization: Maximizing the predicted value of conductivity; Minimizing the total content of nano-silver powder; Satisfying dynamic process constraints: for each process parameter k, calculate the absolute difference between the current formula parameter value and the threshold value, subtract the adaptive tolerance coefficient, take the positive part, and require the value to be not more than zero; Dynamically adjusting the tolerance coefficient: multiplying the initial tolerance coefficient by the adjustment factor raised to the power of t, wherein the adjustment factor is 1 plus the learning rate multiplied by the proportion of individuals violating the constraints in the population size; Binding the process constraints of viscosity range 70-130 Pa・s, fineness index ≤14 microns, and sintering temperature 100-200℃.

4. The machine learning based silver paste conductive performance data modeling and formulation optimization system of claim 1, wherein: The simulation verification module includes: Sintering process simulation unit: simulating the silver-copper alloying process of silver paste in a 100-200℃ sintering environment based on finite element analysis, predicting the interface contact resistance and anti-migration performance; Laser enhanced contact optimization technology adaptation unit: optimize the mass ratio of etherified melamine formaldehyde resin and polyether polyol resin in the formula 25-45%:30-55% to adapt to the laser enhanced contact optimization process of TOPCon cell.

5. The machine learning based silver paste conductive performance data modeling and formulation optimization system of claim 1, wherein: The operations performed by the data preprocessing module include: Imputing missing data: using all decision trees in the random forest to average the prediction results of known feature values as missing feature values; Detecting and deleting outliers identified by isolation forest algorithm; Generating enhanced features: including the mass ratio of nano-silver powder and photovoltaic glass powder, the product of sintering temperature and curing time.

6. The machine learning based silver paste conductive performance data modeling and formulation optimization system of claim 1, wherein: The machine learning model is a double-layer integrated structure: Primary prediction layer: XGBoost regression model to predict electrical conductivity; Random forest classification model to predict weldability; Meta-model layer: using the output of the primary prediction layer as features, support vector machine is used to generate a comprehensive performance score.

7. The machine learning based silver paste conductive performance data modeling and formulation optimization system of claim 3, wherein: The dynamic process constraints include: Viscosity range: 70-130 Pa・s; Fineness index: ≤14 microns; Sintering temperature range: 100-200℃.

8. The machine learning based silver paste conductive performance data modeling and formulation optimization system of claim 1, wherein: The silver paste formula contains the following components and content constraints: Nano-silver powder: 70-80 parts, D50 particle size 0.5-4 microns; Nano-copper powder: 0.1-1 parts, D50 particle size 30-100 nanometers; Photovoltaic glass powder: 1-5 parts, components include 50-60 weight percent lead oxide, 15-25 weight percent silicon dioxide, 10-20 weight percent boron trioxide; Organic carrier: 15-30 parts, including etherified melamine formaldehyde resin and polyether polyol resin.

9. The machine learning based silver paste conductive performance data modeling and formulation optimization system of claim 1, wherein: The silver paste formula optimization method of the system includes: Input target performance constraints: electrical conductivity > 5 x 10 5 Siemens / Miller, weld pull > 3 Newtons / 1.6 millimeters; Identify high-impact components and lock their content candidate set; Generate a set of candidate formulas and screen a set of Pareto optimal solutions through sintering simulation; Experimental verification of the optimized formula and feedback to the data acquisition module.

Images